[
https://issues.apache.org/jira/browse/FLINK-1271?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14222151#comment-14222151
]
Stephan Ewen commented on FLINK-1271:
-------------------------------------
I think we could drop the limitation to {Writable} in the Hadoop formats. The
Writable limitation does not really come from the input format, but only from
the data movement of Hadoop. Since we operate differently anyways, I see no
fundamental reason to restrict the Hadoop Input Format types to Writables.
> Extend HadoopOutputFormat and HadoopInputFormat to handle Void.class
> ---------------------------------------------------------------------
>
> Key: FLINK-1271
> URL: https://issues.apache.org/jira/browse/FLINK-1271
> Project: Flink
> Issue Type: Wish
> Components: Hadoop Compatibility
> Reporter: Felix Neutatz
> Priority: Minor
> Labels: Columnstore, HadoopInputFormat, HadoopOutputFormat,
> Parquet
>
> Parquet, one of the most famous and efficient column store formats in Hadoop
> uses Void.class as Key!
> At the moment there are only keys allowed which extend Writable.
> For example, we would need to be able to do something like:
> HadoopInputFormat hadoopInputFormat = new HadoopInputFormat(new
> ParquetThriftInputFormat(), Void.class, AminoAcid.class, job);
> ParquetThriftInputFormat.addInputPath(job, new Path("newpath"));
> ParquetThriftInputFormat.setReadSupportClass(job, AminoAcid.class);
> // Create a Flink job with it
> DataSet<Tuple2<Void, AminoAcid>> data = env.createInput(hadoopInputFormat);
> Where AminoAcid is a generated Thrift class in this case.
> However, I figured out how to output Parquet files with Parquet by creating a
> class which extends HadoopOutputFormat.
> Now we will have to discuss, what's the best approach to make the Parquet
> integration happen
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
